Robust Self-augmentation for NER with Meta-reweighting

This repository contains the code for Robust Self-Augmentation for Named Entity Recognition with Meta Reweighting (NAACL2022).

Requirements

• Python >= 3.6
• Torch >= 1.3
• transformers >= 4.0
• higher
  • Core Thought: the complex calculation of higher-order gradients is simplified to a first-order approximation (e.g., to do the first-order Taylor expansion)

Prepare

1. Get partial training set: python processing/sample.py 0.05|0.1|0.3
2. Build the entity dictionary: python processing/build_ner_dic.py train_data_file ent.dic cn|en
3. Obtain the word-to-vectors trained on Wikipedia
4. Produce pseudo-labeled training set：python processing/cn|en_aug_util.py train_data_file aug_train_data_file ent.dic ratio aug_times

  Note: The data format is BIOES CoNLL. The processing/conll_util.py script provides the format transformation.

Related Work

• Learning to Reweight Examples for Robust Deep Learning
• Meta-Weight-Net: Learning an Explicit Mapping For Sample Weighting
• Distilling Effective Supervision from Severe Label Noise
• Meta Soft Label Generation for Noisy Labels
• Meta Label Correction for Noisy Label Learning
• Semi-Supervised Learning with Meta-Gradient